Chia1104.dev

The source of chia1104.dev: a personal site, its admin dashboard and the services behind them. It is a pnpm + Turborepo monorepo where new stack and architecture ideas are tried in production.

Architecture

App	Stack and deployment	Responsibility
apps/www	Next.js 16 on Vercel, port 3000	Public profile, blog, projects, search and reading agent
apps/dash	Next.js 16 on Railway, port 3001	Content administration, assets, writing agent and RAG operations
apps/service	Hono on Nitro, Railway, port 3005	Auth, database access, oRPC and AI HTTP routes
apps/workflow	Hono on Nitro, Railway, port 3008	Durable workflows, indexing jobs and agent turns; one replica

Both frontends call service through a contract-first oRPC client. service sends durable work to workflow through @chia/workflow-control; both read the same PostgreSQL-backed Workflow World.

The main infrastructure is Better Auth, Postgres with Drizzle, Redis through Keyv, the Vercel Workflow SDK, and ParadeDB for BM25 plus vector retrieval. Content is authored as MDX and rendered with Fumadocs.

Back-office AI features

Writing agent

The dashboard includes a multi-session writing workspace. The agent can read existing posts, search and fetch web sources, keep long-term memory, prepare locale-specific drafts and commit approved changes to live content. Sessions support streaming, abort, compaction, rewind and forks; commit-tier tools pass through the approval policy before writing live content.

The administration pages also provide:

• Agent-kind defaults and instructions.
• Model, prompt and parameter overrides for title, compaction, branch-summary and lesson-extraction tasks.
• Weekly house-spend and concurrent-turn limits for visitors.
• Memory review for fetched sources, saved facts and learned writing preferences, including approval of pending lessons and manual session consolidation.

The public site has a separate read-only agent with published-content access and a smaller model and tool budget.

See Agent architecture or the 中文版本.

RAG management

The RAG dashboard manages the shared chunk and embedding index used by search and agent memory:

• Overview of the active model and index version, embedding coverage, stale or missing vectors, and breakdowns by source, locale, chunk kind and visibility.
• Chunk explorer with content, embedding-state, kind and locale filters, plus per-chunk details.
• Index-run history with target, progress, failures and duration.
• Maintenance actions to embed missing chunks, run a full reindex and prune stale vectors.

See RAG architecture or the 中文版本.

Repository layout

apps/          deployable applications and scheduled functions
packages/      source-exported @chia/* packages
tests/         Playwright end-to-end tests
docs/          current agent, RAG and workflow architecture
toolings/      shared configuration, scripts and generators
infra/railway/ Railway service configuration
legacy/        reference-only code; never imported

Development rules and package boundaries live in AGENTS.md, apps/AGENTS.md and packages/AGENTS.md.

Getting started

Requirements: Node.js 22 or newer, pnpm 12.1.0 and Docker. Pkl 0.25.2 or newer is needed only when regenerating site metadata.

git clone https://github.com/chia1104/chia1104.dev.git
cd chia1104.dev
make init
pnpm db:up
pnpm db:migrate
pnpm db:seed # optional

make init installs dependencies and copies .env.example to .env for www, dash, service and workflow. Fill in the required values before starting the apps.

Commands

pnpm dev:www          # www + service + workflow
pnpm dev:dash         # dash + service + workflow
pnpm dev:service      # service + workflow
pnpm dev              # all workspaces

pnpm build            # build all workspaces
pnpm build:www        # or build:dash, build:service, build:workflow
pnpm lint
pnpm format
pnpm type:check
pnpm test
pnpm test:e2e

pnpm db:generate
pnpm db:migrate
pnpm db:push
pnpm db:studio

make help lists equivalent setup, development, validation, database and Docker targets.

Deployment

www deploys to Vercel. dash, service and the single-replica workflow deploy to Railway; their settings live in infra/railway/. Run make docker-build to build all local images or use the per-app Docker targets listed by make help.

The workflow runner must remain single-replica until it has cross-process replay and step claims. See Workflow deployment.

License

MIT. See LICENSE.